TCMDATA: an integrated R package for target-oriented network pharmacology analysis and AI-assisted interpretation in traditional Chinese medicine

• TCMDATA is a comprehensive R toolkit for Traditional Chinese Medicine (TCM) network pharmacology research.
• It provides an end-to-end computational workflow—from herb–compound–target data retrieval and pharmacological network construction, through enrichment analysis and PPI mining, to machine-learning-based biomarker screening and AI-powered result interpretation.
• Publication-ready visualizations including Sankey diagrams, docking heatmaps, lollipop plots, and more.

For details, please visit the full documentation.

---

Highlights

• 🌿 Built-in TCM Database: Manually curated herb–compound–target interaction data ready for analysis
• 🔬 PubChem Integration: Compound identification, property retrieval, and structure download
• 📊 PPI Network Analysis: STRING-based network retrieval, 17+ centrality metrics, community detection, robustness evaluation, and weighted disease-module proximity analysis for target or formula prioritization
• 🤖 Machine Learning-based Feature Selection: 6 algorithms (LASSO, Elastic Net, Ridge, RF+Boruta, SVM-RFE, XGBoost) with consensus scoring
• 🧠 AI Agent System: LLM-powered agent with natural language task routing, skill-based workflow, and interactive chat (tcm_agent(), tcm_chat())
• 📚 Literature & Data Mining: PubMed search and GEO dataset discovery for TCM–disease studies
• 🔗 Seamless Integration: Works with clusterProfiler, enrichplot, and other Bioconductor tools for enrichment analysis

Feature Overview

Module	Description	Key Functions
Data Retrieval	Bidirectional query of herbs, compounds, and validated targets	search_herb(), search_target()
Disease Targets	Open Targets disease–target queries and GEO dataset discovery	query_disease_targets(), search_disease_efo(), search_geo_datasets()
Molecule Detection	PubChem-based CID resolution, property annotation, similarity search	resolve_cid(), getprops(), compound_similarity()
Network Construction	Build herb–compound–target networks with topological metrics	prepare_herb_graph()
Enrichment Analysis	Herb-based over-representation analysis; GO/KEGG compatible	herb_enricher()
PPI Analysis	STRING retrieval, 17+ centrality metrics, community detection, robustness, and weighted target-set proximity to disease modules	get_ppi(), compute_nodeinfo(), rank_tcm_targets_by_ppi(), evaluate_tcm_target_sets_by_ppi(), ppi_knock()
Clustering	Louvain, MCL, and MCODE community detection	run_louvain(), run_MCL(), run_mcode()
ML Screening	6 algorithms × 3 validation modes with consensus analysis	run_ml_screening(), plot_ml_roc()
AI Agent	Natural language task routing, skill-based workflow, interactive chat	tcm_agent(), tcm_chat(), tcm_interpret()
AI Interpretation	LLM-powered interpretation for enrichment, PPI, tables	tcm_interpret(), draft_result_paragraph()
Visualization	Sankey, docking heatmaps, lollipop, UpSet, radar charts	tcm_sankey(), ggdock(), gglollipop(), upsetplot()

Installation

# install.packages("remotes")
options(timeout = 600)
remotes::install_github("YuLab-SMU/TCMDATA")

Quick Start

library(TCMDATA)

# Search by herb name (supports Chinese pinyin)
huangqi <- search_herb("huangqi", "Herb_pinyin_name")
head(huangqi)

# Reverse lookup: find herbs targeting a specific gene
il6_herbs <- search_target("IL6")
head(il6_herbs)

AI Agent System

TCMDATA integrates an AI agent layer via aisdk, featuring natural language task routing, a skill-based workflow engine, and 36 registered tools.

# One-time setup
remotes::install_github("YuLab-SMU/aisdk")

# Also required when using DeepSeek or another additional provider
remotes::install_github("YuLab-SMU/aisdk.providers")

tcm_setup(
  provider = "openai",
  api_key  = "sk-xxxx",
  model    = "gpt-6.1-sol",
  save     = TRUE
)

# DeepSeek uses the companion provider package
tcm_setup("deepseek", "sk-xxxx", "deepseek-chat")

# OpenAI-compatible relay endpoint
tcm_setup(
  provider = "anthropic",
  api_key = "relay-key",
  model = "relay-claude-model-id",
  base_url = "https://relay.example.com/v1",
  api_format = "chat_completions"
)

# One-shot agent: auto-routes task to appropriate tools
result <- tcm_agent("Query the targets of Astragalus and list the top 10 most frequently occurring ones.")

# Interactive multi-turn chat session
tcm_chat()

# AI-powered result interpretation
ai_res <- tcm_interpret(enrich_res, language = "en")
draft  <- draft_result_paragraph(ai_res, language = "en")

The agent supports 11 task types including herb/disease/target lookup, enrichment, PPI analysis, ML screening, PubMed search, GEO dataset discovery, and more. Plot requests create artifacts that are automatically exported to .GlobalEnv. Each tool-using chat turn also exports a replayable script such as tcm_script_001 for inspection and reuse. Skills can be customized via tcm_init_skills().

Documentation

Complete tutorials with worked examples can be found here.

Citation

If you use TCMDATA in your research, please cite:

@article{he2026tcmdata,
  title   = {TCMDATA: An integrated R package for target-oriented network pharmacology analysis and AI-assisted interpretation in traditional Chinese medicine},
  author  = {He, Nan and Xu, Shuangbin and Mo, Ke and Xia, Yonghe and Wang, Rui and Tao, Shaoguo and Wang, Qianwen and Mo, Jianwen and Pang, Bin and Qiu, Yiyan and Song, Qiong and Yu, Guangchuang},
  journal = {Current Molecular Pharmacology},
  volume  = {19},
  number  = {1},
  pages   = {372--388},
  year    = {2026},
  issn    = {1874-4672},
  doi     = {10.1016/j.cmp.2026.09.003},
  url     = {https://doi.org/10.1016/j.cmp.2026.09.003}
}